In the field of flow sensing, there are trends towards the realization of complex functional systems that are able to extract many different fluid parameters rather than individual components that provide a turn-key solution to a flow-problem. Microfabrication technologies support this trend by enabling the integration of multiple sensors - e.g. flow rate, pressure and density sensors - into one chip.
The main challenge is that, in thermal flow sensors, the output signal of the flow meter is not only dependent on the flow rate, but also on thermal physical properties of the fluid. This requires a thorough characterization and calibration of a combination of multiple non-ideal sensing structures. However, the calibration is tedious and costly.
This project aims at exploiting compressive deep learning models such that the sensor system will be able to learn and predict fluid flow rate without a calibration.
The project will be divided into 3 main tasks:
- Study flow rate estimation with compressed and non-compressed deep neural networks
- Compare the networks with regard the flow rate prediction problem.
- Implement the most suitable compressed model on Pi Zero.
30% Theory, 50% Implementation, 20%Writing
Le Viet Duc, email@example.com, room ZI 5013